Merge pull request #1703 from emankov/HIPIFY

[HIPIFY][rocBLAS] 64-bit functions support - Step 19
ROCm · Oct 13, 2024 · 1c16c89 · 1c16c89
2 parents f630919 + 8719384
commit 1c16c89
Show file tree

Hide file tree

Showing 5 changed files with 65 additions and 32 deletions.
diff --git a/bin/hipify-perl b/bin/hipify-perl
@@ -1714,7 +1714,9 @@ sub rocSubstitutions {
     subst("cublasCtrmv_v2_64", "rocblas_ctrmv_64", "library");
     subst("cublasCtrsm", "rocblas_ctrsm", "library");
     subst("cublasCtrsmBatched", "rocblas_ctrsm_batched", "library");
+    subst("cublasCtrsm_64", "rocblas_ctrsm_64", "library");
     subst("cublasCtrsm_v2", "rocblas_ctrsm", "library");
+    subst("cublasCtrsm_v2_64", "rocblas_ctrsm_64", "library");
     subst("cublasCtrsv", "rocblas_ctrsv", "library");
     subst("cublasCtrsv_64", "rocblas_ctrsv_64", "library");
     subst("cublasCtrsv_v2", "rocblas_ctrsv", "library");
@@ -1846,7 +1848,9 @@ sub rocSubstitutions {
     subst("cublasDtrmv_v2_64", "rocblas_dtrmv_64", "library");
     subst("cublasDtrsm", "rocblas_dtrsm", "library");
     subst("cublasDtrsmBatched", "rocblas_dtrsm_batched", "library");
+    subst("cublasDtrsm_64", "rocblas_dtrsm_64", "library");
     subst("cublasDtrsm_v2", "rocblas_dtrsm", "library");
+    subst("cublasDtrsm_v2_64", "rocblas_dtrsm_64", "library");
     subst("cublasDtrsv", "rocblas_dtrsv", "library");
     subst("cublasDtrsv_64", "rocblas_dtrsv_64", "library");
     subst("cublasDtrsv_v2", "rocblas_dtrsv", "library");
@@ -2062,7 +2066,9 @@ sub rocSubstitutions {
     subst("cublasStrmv_v2_64", "rocblas_strmv_64", "library");
     subst("cublasStrsm", "rocblas_strsm", "library");
     subst("cublasStrsmBatched", "rocblas_strsm_batched", "library");
+    subst("cublasStrsm_64", "rocblas_strsm_64", "library");
     subst("cublasStrsm_v2", "rocblas_strsm", "library");
+    subst("cublasStrsm_v2_64", "rocblas_strsm_64", "library");
     subst("cublasStrsv", "rocblas_strsv", "library");
     subst("cublasStrsv_64", "rocblas_strsv_64", "library");
     subst("cublasStrsv_v2", "rocblas_strsv", "library");
@@ -2217,7 +2223,9 @@ sub rocSubstitutions {
     subst("cublasZtrmv_v2_64", "rocblas_ztrmv_64", "library");
     subst("cublasZtrsm", "rocblas_ztrsm", "library");
     subst("cublasZtrsmBatched", "rocblas_ztrsm_batched", "library");
+    subst("cublasZtrsm_64", "rocblas_ztrsm_64", "library");
     subst("cublasZtrsm_v2", "rocblas_ztrsm", "library");
+    subst("cublasZtrsm_v2_64", "rocblas_ztrsm_64", "library");
     subst("cublasZtrsv", "rocblas_ztrsv", "library");
     subst("cublasZtrsv_64", "rocblas_ztrsv_64", "library");
     subst("cublasZtrsv_v2", "rocblas_ztrsv", "library");
@@ -12675,8 +12683,6 @@ sub warnRocOnlyUnsupportedFunctions {
     my $k = 0;
     foreach $func (
         "cublasZtrttp",
-        "cublasZtrsm_v2_64",
-        "cublasZtrsm_64",
         "cublasZtrsmBatched_64",
         "cublasZtrmm_v2_64",
         "cublasZtrmm_64",
@@ -12714,8 +12720,6 @@ sub warnRocOnlyUnsupportedFunctions {
         "cublasSwapEx_64",
         "cublasSwapEx",
         "cublasStrttp",
-        "cublasStrsm_v2_64",
-        "cublasStrsm_64",
         "cublasStrsmBatched_64",
         "cublasStrmm_v2_64",
         "cublasStrmm_64",
@@ -12847,8 +12851,6 @@ sub warnRocOnlyUnsupportedFunctions {
         "cublasGemmBatchedEx_64",
         "cublasFree",
         "cublasDtrttp",
-        "cublasDtrsm_v2_64",
-        "cublasDtrsm_64",
         "cublasDtrsmBatched_64",
         "cublasDtrmm_v2_64",
         "cublasDtrmm_64",
@@ -12875,8 +12877,6 @@ sub warnRocOnlyUnsupportedFunctions {
         "cublasDgeam_64",
         "cublasDdgmm_64",
         "cublasCtrttp",
-        "cublasCtrsm_v2_64",
-        "cublasCtrsm_64",
         "cublasCtrsmBatched_64",
         "cublasCtrmm_v2_64",
         "cublasCtrmm_64",

diff --git a/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md b/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md
@@ -1068,9 +1068,9 @@
 |`cublasCtrmm_v2`| | | | |`hipblasCtrmm_v2`|6.0.0| | | | |`rocblas_ctrmm`|3.5.0| |6.0.0| | |
 |`cublasCtrmm_v2_64`|12.0| | | | | | | | | | | | | | | |
 |`cublasCtrsm`| | | | |`hipblasCtrsm_v2`|6.0.0| | | | |`rocblas_ctrsm`|3.5.0| | | | |
-|`cublasCtrsm_64`|12.0| | | | | | | | | | | | | | | |
+|`cublasCtrsm_64`|12.0| | | | | | | | | |`rocblas_ctrsm_64`|6.2.0| | | | |
 |`cublasCtrsm_v2`| | | | |`hipblasCtrsm_v2`|6.0.0| | | | |`rocblas_ctrsm`|3.5.0| | | | |
-|`cublasCtrsm_v2_64`|12.0| | | | | | | | | | | | | | | |
+|`cublasCtrsm_v2_64`|12.0| | | | | | | | | |`rocblas_ctrsm_64`|6.2.0| | | | |
 |`cublasDgemm`| | | | |`hipblasDgemm`|1.8.2| | | | |`rocblas_dgemm`|1.5.0| | | | |
 |`cublasDgemmBatched`| | | | |`hipblasDgemmBatched`|1.8.2| | | | |`rocblas_dgemm_batched`|3.5.0| | | | |
 |`cublasDgemmBatched_64`|12.0| | | | | | | | | | | | | | | |
@@ -1104,9 +1104,9 @@
 |`cublasDtrmm_v2`| | | | |`hipblasDtrmm`|3.2.0| |6.0.0| | |`rocblas_dtrmm`|3.5.0| |6.0.0| | |
 |`cublasDtrmm_v2_64`|12.0| | | | | | | | | | | | | | | |
 |`cublasDtrsm`| | | | |`hipblasDtrsm`|1.8.2| | | | |`rocblas_dtrsm`|1.5.0| | | | |
-|`cublasDtrsm_64`|12.0| | | | | | | | | | | | | | | |
+|`cublasDtrsm_64`|12.0| | | | | | | | | |`rocblas_dtrsm_64`|6.2.0| | | | |
 |`cublasDtrsm_v2`| | | | |`hipblasDtrsm`|1.8.2| | | | |`rocblas_dtrsm`|1.5.0| | | | |
-|`cublasDtrsm_v2_64`|12.0| | | | | | | | | | | | | | | |
+|`cublasDtrsm_v2_64`|12.0| | | | | | | | | |`rocblas_dtrsm_64`|6.2.0| | | | |
 |`cublasGemmGroupedBatchedEx`|12.5| | | | | | | | | | | | | | | |
 |`cublasGemmGroupedBatchedEx_64`|12.5| | | | | | | | | | | | | | | |
 |`cublasHSHgemvBatched`|11.6| | | | | | | | | |`rocblas_hshgemv_batched`|6.0.0| | | | |
@@ -1156,9 +1156,9 @@
 |`cublasStrmm_v2`| | | | |`hipblasStrmm`|3.2.0| |6.0.0| | |`rocblas_strmm`|3.5.0| |6.0.0| | |
 |`cublasStrmm_v2_64`|12.0| | | | | | | | | | | | | | | |
 |`cublasStrsm`| | | | |`hipblasStrsm`|1.8.2| | | | |`rocblas_strsm`|1.5.0| | | | |
-|`cublasStrsm_64`|12.0| | | | | | | | | | | | | | | |
+|`cublasStrsm_64`|12.0| | | | | | | | | |`rocblas_strsm_64`|6.2.0| | | | |
 |`cublasStrsm_v2`| | | | |`hipblasStrsm`|1.8.2| | | | |`rocblas_strsm`|1.5.0| | | | |
-|`cublasStrsm_v2_64`|12.0| | | | | | | | | | | | | | | |
+|`cublasStrsm_v2_64`|12.0| | | | | | | | | |`rocblas_strsm_64`|6.2.0| | | | |
 |`cublasTSSgemvBatched`|11.6| | | | | | | | | |`rocblas_tssgemv_batched`|6.0.0| | | | |
 |`cublasTSSgemvBatched_64`|12.0| | | | | | | | | |`rocblas_tssgemv_batched_64`|6.2.0| | | | |
 |`cublasTSSgemvStridedBatched`|11.6| | | | | | | | | |`rocblas_tssgemv_strided_batched`|6.0.0| | | | |
@@ -1214,9 +1214,9 @@
 |`cublasZtrmm_v2`| | | | |`hipblasZtrmm_v2`|6.0.0| | | | |`rocblas_ztrmm`|3.5.0| |6.0.0| | |
 |`cublasZtrmm_v2_64`|12.0| | | | | | | | | | | | | | | |
 |`cublasZtrsm`| | | | |`hipblasZtrsm_v2`|6.0.0| | | | |`rocblas_ztrsm`|3.5.0| | | | |
-|`cublasZtrsm_64`|12.0| | | | | | | | | | | | | | | |
+|`cublasZtrsm_64`|12.0| | | | | | | | | |`rocblas_ztrsm_64`|6.2.0| | | | |
 |`cublasZtrsm_v2`| | | | |`hipblasZtrsm_v2`|6.0.0| | | | |`rocblas_ztrsm`|3.5.0| | | | |
-|`cublasZtrsm_v2_64`|12.0| | | | | | | | | | | | | | | |
+|`cublasZtrsm_v2_64`|12.0| | | | | | | | | |`rocblas_ztrsm_64`|6.2.0| | | | |
 
 ## **8. BLAS-like Extension**
 

diff --git a/docs/tables/CUBLAS_API_supported_by_ROC.md b/docs/tables/CUBLAS_API_supported_by_ROC.md
@@ -1068,9 +1068,9 @@
 |`cublasCtrmm_v2`| | | | |`rocblas_ctrmm`|3.5.0| |6.0.0| | |
 |`cublasCtrmm_v2_64`|12.0| | | | | | | | | |
 |`cublasCtrsm`| | | | |`rocblas_ctrsm`|3.5.0| | | | |
-|`cublasCtrsm_64`|12.0| | | | | | | | | |
+|`cublasCtrsm_64`|12.0| | | |`rocblas_ctrsm_64`|6.2.0| | | | |
 |`cublasCtrsm_v2`| | | | |`rocblas_ctrsm`|3.5.0| | | | |
-|`cublasCtrsm_v2_64`|12.0| | | | | | | | | |
+|`cublasCtrsm_v2_64`|12.0| | | |`rocblas_ctrsm_64`|6.2.0| | | | |
 |`cublasDgemm`| | | | |`rocblas_dgemm`|1.5.0| | | | |
 |`cublasDgemmBatched`| | | | |`rocblas_dgemm_batched`|3.5.0| | | | |
 |`cublasDgemmBatched_64`|12.0| | | | | | | | | |
@@ -1104,9 +1104,9 @@
 |`cublasDtrmm_v2`| | | | |`rocblas_dtrmm`|3.5.0| |6.0.0| | |
 |`cublasDtrmm_v2_64`|12.0| | | | | | | | | |
 |`cublasDtrsm`| | | | |`rocblas_dtrsm`|1.5.0| | | | |
-|`cublasDtrsm_64`|12.0| | | | | | | | | |
+|`cublasDtrsm_64`|12.0| | | |`rocblas_dtrsm_64`|6.2.0| | | | |
 |`cublasDtrsm_v2`| | | | |`rocblas_dtrsm`|1.5.0| | | | |
-|`cublasDtrsm_v2_64`|12.0| | | | | | | | | |
+|`cublasDtrsm_v2_64`|12.0| | | |`rocblas_dtrsm_64`|6.2.0| | | | |
 |`cublasGemmGroupedBatchedEx`|12.5| | | | | | | | | |
 |`cublasGemmGroupedBatchedEx_64`|12.5| | | | | | | | | |
 |`cublasHSHgemvBatched`|11.6| | | |`rocblas_hshgemv_batched`|6.0.0| | | | |
@@ -1156,9 +1156,9 @@
 |`cublasStrmm_v2`| | | | |`rocblas_strmm`|3.5.0| |6.0.0| | |
 |`cublasStrmm_v2_64`|12.0| | | | | | | | | |
 |`cublasStrsm`| | | | |`rocblas_strsm`|1.5.0| | | | |
-|`cublasStrsm_64`|12.0| | | | | | | | | |
+|`cublasStrsm_64`|12.0| | | |`rocblas_strsm_64`|6.2.0| | | | |
 |`cublasStrsm_v2`| | | | |`rocblas_strsm`|1.5.0| | | | |
-|`cublasStrsm_v2_64`|12.0| | | | | | | | | |
+|`cublasStrsm_v2_64`|12.0| | | |`rocblas_strsm_64`|6.2.0| | | | |
 |`cublasTSSgemvBatched`|11.6| | | |`rocblas_tssgemv_batched`|6.0.0| | | | |
 |`cublasTSSgemvBatched_64`|12.0| | | |`rocblas_tssgemv_batched_64`|6.2.0| | | | |
 |`cublasTSSgemvStridedBatched`|11.6| | | |`rocblas_tssgemv_strided_batched`|6.0.0| | | | |
@@ -1214,9 +1214,9 @@
 |`cublasZtrmm_v2`| | | | |`rocblas_ztrmm`|3.5.0| |6.0.0| | |
 |`cublasZtrmm_v2_64`|12.0| | | | | | | | | |
 |`cublasZtrsm`| | | | |`rocblas_ztrsm`|3.5.0| | | | |
-|`cublasZtrsm_64`|12.0| | | | | | | | | |
+|`cublasZtrsm_64`|12.0| | | |`rocblas_ztrsm_64`|6.2.0| | | | |
 |`cublasZtrsm_v2`| | | | |`rocblas_ztrsm`|3.5.0| | | | |
-|`cublasZtrsm_v2_64`|12.0| | | | | | | | | |
+|`cublasZtrsm_v2_64`|12.0| | | |`rocblas_ztrsm_64`|6.2.0| | | | |
 
 ## **8. BLAS-like Extension**
 

diff --git a/src/CUDA2HIP_BLAS_API_functions.cpp b/src/CUDA2HIP_BLAS_API_functions.cpp
@@ -541,13 +541,13 @@ const std::map<llvm::StringRef, hipCounter> CUDA_BLAS_FUNCTION_MAP {
 
   // TRSM
   {"cublasStrsm",                                          {"hipblasStrsm",                                              "rocblas_strsm",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_SUPPORTED_V2_ONLY}},
-  {"cublasStrsm_64",                                       {"hipblasStrsm_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
+  {"cublasStrsm_64",                                       {"hipblasStrsm_64",                                           "rocblas_strsm_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_UNSUPPORTED}},
   {"cublasDtrsm",                                          {"hipblasDtrsm",                                              "rocblas_dtrsm",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_SUPPORTED_V2_ONLY}},
-  {"cublasDtrsm_64",                                       {"hipblasDtrsm_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
+  {"cublasDtrsm_64",                                       {"hipblasDtrsm_64",                                           "rocblas_dtrsm_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_UNSUPPORTED}},
   {"cublasCtrsm",                                          {"hipblasCtrsm_v2",                                           "rocblas_ctrsm",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_SUPPORTED_V2_ONLY}},
-  {"cublasCtrsm_64",                                       {"hipblasCtrsm_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
+  {"cublasCtrsm_64",                                       {"hipblasCtrsm_64",                                           "rocblas_ctrsm_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_UNSUPPORTED}},
   {"cublasZtrsm",                                          {"hipblasZtrsm_v2",                                           "rocblas_ztrsm",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_SUPPORTED_V2_ONLY}},
-  {"cublasZtrsm_64",                                       {"hipblasZtrsm_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
+  {"cublasZtrsm_64",                                       {"hipblasZtrsm_64",                                           "rocblas_ztrsm_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_UNSUPPORTED}},
 
   // TRMM
   {"cublasStrmm",                                          {"hipblasStrmm",                                              "rocblas_strmm",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_SUPPORTED_V2_ONLY}},
@@ -908,13 +908,13 @@ const std::map<llvm::StringRef, hipCounter> CUDA_BLAS_FUNCTION_MAP {
 
   // TRSM
   {"cublasStrsm_v2",                                       {"hipblasStrsm",                                              "rocblas_strsm",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3}},
-  {"cublasStrsm_v2_64",                                    {"hipblasStrsm_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
+  {"cublasStrsm_v2_64",                                    {"hipblasStrsm_64",                                           "rocblas_strsm_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_UNSUPPORTED}},
   {"cublasDtrsm_v2",                                       {"hipblasDtrsm",                                              "rocblas_dtrsm",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3}},
-  {"cublasDtrsm_v2_64",                                    {"hipblasDtrsm_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
+  {"cublasDtrsm_v2_64",                                    {"hipblasDtrsm_64",                                           "rocblas_dtrsm_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_UNSUPPORTED}},
   {"cublasCtrsm_v2",                                       {"hipblasCtrsm_v2",                                           "rocblas_ctrsm",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3}},
-  {"cublasCtrsm_v2_64",                                    {"hipblasCtrsm_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
+  {"cublasCtrsm_v2_64",                                    {"hipblasCtrsm_64",                                           "rocblas_ctrsm_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_UNSUPPORTED}},
   {"cublasZtrsm_v2",                                       {"hipblasZtrsm_v2",                                           "rocblas_ztrsm",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3}},
-  {"cublasZtrsm_v2_64",                                    {"hipblasZtrsm_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
+  {"cublasZtrsm_v2_64",                                    {"hipblasZtrsm_64",                                           "rocblas_ztrsm_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_UNSUPPORTED}},
 
   // TRMM
   {"cublasStrmm_v2",                                       {"hipblasStrmm",                                              "rocblas_strmm",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3}},
@@ -2403,6 +2403,10 @@ const std::map<llvm::StringRef, hipAPIversions> HIP_BLAS_FUNCTION_VER_MAP {
   {"rocblas_zgeru_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
   {"rocblas_cgerc_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
   {"rocblas_zgerc_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
+  {"rocblas_strsm_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
+  {"rocblas_dtrsm_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
+  {"rocblas_ctrsm_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
+  {"rocblas_ztrsm_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
 };
 
 const std::map<llvm::StringRef, hipAPIChangedVersions> HIP_BLAS_FUNCTION_CHANGED_VER_MAP {